Abstract:
The Virtual World Data Server: A Scalable Multimedia Object Server
https://dml.cs.ucla.edu/vwds/

Multimedia information systems have become the enabling component for a number of existing application domains such as Video-Conferencing and "Video On Demand" services, as well as entirely new ones. For example, Interactive Virtual Reality can model anything from modern cities to even imaginary worlds of electric and magnetic fields normally invisible to the naked eye, while at the same time providing an inherently visual interface for knowledge discovery. Within the next few years, the increased demand for these existing and next generation applications will continue to highlight the inadequacies of conventional multimedia systems.

For example, the resource requirements of Interactive Virtual Reality models are potentially huge, with a realistic model of the Los Angeles basin containing over 50 gigabytes of texture alone. Conventional "multimedia Servers" are inadequate for delivering 3D virtual reality because the periodic and predictable delivery optimizations of video streams are the antithesis to the seemingly random and unpredictable movements in a 3D world. A server that efficiently stores and serves such large data sets while meeting real-time delivery constraints will require novel system designs and implementations.

The UCLA Multimedia Laboratory has designed and built such a next-generation multimedia server called the "Virtual World Data Server" (VWDS) using a Sun Ultra Enterprise E4000 with 10 UltraSparc processors, 1.25 GB of main memory and 56 GB of raw storage. This monolithic solution provides a powerful facility for the implementation of VWDS, but is inherently limited in scale. Therefore, we have begun the investigation phase of designing and implementing VWDS on a cluster of commodity PCs running Microsoft's Windows NT operating system.

Moving from an SMP to a shared nothing architecture adds a number of interesting research challenges to those of building an Interactive Virtual Reality system:

- How can we approach the optimizations and flexibilities that an SMP architecture provides in a clustered environment? What are the tradeoffs?

- How have the performance bottlenecks changed for meeting the real-time delivery requirements?

- What are the inherent scaling limitations of such a system? How can such limitations be addressed without compromising performance?

- How can newer technologies such as Fibre Channel be used to improve performance and the systems ability to scale?

- Is it possible to extend the system to a fully distributed one? What are the limitations of such a system?

- How must the clients of the system change? Will they be required to do more than in the monolithic solution or less?

- How well can the system be used as a backbone for other application domains such as collaborative interactions?
- Finally, will Windows NT provide an adequate foundation upon which to build such a system?

Frank Fabbrocino